---
license: apache-2.0
library_name: diffusers
pipeline_tag: image-to-video
tags:
  - wan
  - wan2.2
  - image-to-video
  - video-generation
  - cineforge
  - fp8
---

# CineForge Wan Models

This repository is the model distribution channel for CineForge, a standalone local Windows application for Wan video generation.

## Publication status

**Native generation validated; Desktop 0.5.0 release-candidate pack.**

The four core Wan 2.2 I2V A14B scaled-FP8 components are published with provenance, exact byte sizes, and SHA-256 checksums. CineForge's standalone native loader and a real high/low-expert I2V generation were validated on 2026-08-14 without ComfyUI. The repository also includes the pinned scheduler, tokenizer, and architecture configs required by the Desktop loader.

## Intended pack contents

- Wan 2.2 I2V A14B high-noise expert
- Wan 2.2 I2V A14B low-noise expert
- Wan-compatible UMT5 text encoder
- Wan VAE
- CineForge pack manifest and checksums
- upstream and derivative license/notice files
- reproducible conversion and validation records

Optional third-party acceleration LoRAs are excluded from the core pack until their independent origin and license are documented.

## Upstream and provenance

- Wan 2.2 project: https://github.com/Wan-Video/Wan2.2
- Official Wan models: https://huggingface.co/Wan-AI
- Split FP8 source candidates used on the development workstation:
  - https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged
  - https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged

Wan 2.2 is published by the Wan Team under Apache 2.0. Any repackaged or converted artifact must retain the applicable notices and must be verified independently before redistribution.

## Compatibility

| Pack | CineForge | State | Notes |
|---|---|---|---|
| Wan 2.2 I2V A14B scaled-FP8 | 0.5.0 | Release candidate | Native load, two-expert generation, live step telemetry, finite decoded frames, MP4 export, and Desktop installer auto-download flow verified on RTX 4070. Wider hardware validation remains. |

The word `supported` is reserved for packs that load without ComfyUI, complete deterministic generation, provide live progress telemetry, export a decodable video, and pass a clean-machine installation test.

## Application

CineForge source and releases: https://github.com/thebalddudeco/CineForge

The public CineForge Desktop installer automatically downloads this pinned model pack into the user's selected **CineForge Library**. Users do not browse for individual Wan files manually; setup resumes interrupted transfers and verifies every component by file size and SHA-256 before the pack is accepted.

## Independence notice

CineForge is an independent project and is not affiliated with or endorsed by Alibaba, the Wan Team, Hugging Face, or Comfy Org.